Few-shot prompting for consistency

When instructions have already failed

Few-shot examples are the most effective technique for getting consistently formatted output that a person or program can use once detailed instructions have not delivered it. The signal to switch is not complexity — it is inconsistency across runs on similar input.

Two to four, chosen for ambiguity

The useful count is small: two to four targeted examples. General prompting guidance suggests three to five; the difference does not matter much, and what both agree on is that the number is small and the selection is what counts. Choose examples from the ambiguous cases, not the obvious ones. Common cases are usually already handled correctly, so repeating them adds little. Use the space where the model must make a difficult choice, and show why one action was chosen over a plausible alternative.

An example that shows only input and output teaches one mapping. An explanation of why the output is correct teaches the decision rule that can be applied to a new case.

Generalization, not matching

Well-chosen examples teach a decision rule that can apply to new cases. They do more than show the exact inputs they contain. That's why four examples near a difficult boundary can teach more than twenty repetitive examples.

Distinguishing acceptable patterns from genuine issues is the same technique aimed at false positives: show one of each, with the reasoning, and precision improves without a rule for every case.

The following examples show acceptable and unacceptable handling of the same kind of error:

Input:  catch (err) { logger.warn(err); return cached; }
Output: skip
Why:    the failure is handled and the fallback is deliberate

Input:  catch (err) { }
Output: flag
Why:    the error is discarded, so the failure is invisible at runtime

The Why: lines are the part that generalizes. A third case neither example contains — a catch that only increments a counter — is decided by the same question: is the failure still visible afterward?

Extraction and hallucination

Examples are also the lever for reducing hallucination in extraction. Show correct handling of the messy realities — informal measurements, inline citations versus bibliographies, methodology stated up front versus buried in a paragraph.

If a required field comes back empty or invented, show correct extraction from several document structures. The examples teach the model how the same field can appear in different layouts.